NCF
1 means item i is relevant to u, and 0 otherwise
目标函数:用户和app的匹配程度
Wide & Deep Model – ranking model (Google Play,基于用户query,推荐合适的item)
memorization(记忆):wide(LR),线性模型 + 交叉特征(AND)
generalization(泛化):deep(DNN),几乎不需要人工特征工程
ex.
query = “fried chicken”, ->(0,0)
item = “chicken fried rice” -> (0, 1)
word1 word2 wordX word 4
wordX -> word1 word2 word4
Candidate Generation – 超大规模分类问题(类似skip-gram,采用方法:hierarchical softmax 和 negative sample)
youtube把召回阶段建模为一个分类模型,其目标是根据上下文C,用户U,从集合V中找到时刻t最可能被观看的视频Wt
u代表用户和conte